Co-Authoring With AI Users: The Disclosure Question Nobody Wants to Ask

A 2025 University of Kentucky survey of biomedical investigators found that nearly half of researchers reported AI use was never discussed within their teams before work began. At the same time, 42% of those researchers used AI weekly or daily on collaborative projects. The gap between what teams use and what teams talk about sits at the center of academic publishing in 2026.

Most AI policy discussions assume a single author working alone. The reality is different. Almost no research paper has one author. When three of five co-authors used AI to draft different sections, the team faces a chain of unanswered problems: deciding who is responsible for disclosure, deciding whose job verification falls to, and deciding what to do if one author refuses to declare. These problems rarely make it into policy guidance, but they affect every collaborative manuscript headed for submission.

Why the Collaborative Case is Harder Than the Solo Case

Single-author AI disclosure is straightforward. The author knows what tools were used, where, and for what purpose. Disclosure becomes a matter of memory and intention.

Collaborative manuscripts compress the same chain across multiple people, time zones, and disciplines. For example, a statistician may use AI to generate code comments, a clinician co-author uses it to refine the discussion section, and a graduate student uses it for literature summaries. By the time the manuscript reaches the corresponding author, no single person knows what AI touched the document, when, or for what task.

This is the structural problem behind the disclosure failure. The Kentucky study documented widespread concerns about misinformation, bias, and overreliance on AI. The same study also found that decisions about AI use rarely happen at the team level. AI integration is happening individually. Accountability is assigned collectively.

What Current Authorship Guidelines Say About Shared Responsibility

The International Committee of Medical Journal Editors (ICMJE) is direct on this point. Every co-author must take accountability for the integrity of the work, including any sections produced with AI assistance. The corresponding author has specific duties around ensuring "all the journal's administrative requirements, such as providing details of authorship... and disclosures of relationships and activities are properly completed and reported."

Elsevier's policy goes further: "Each (co-)author is accountable for ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved." Frontiers requires that "all authors are responsible for verifying the contributions of all co-authors listed on the manuscript before submission."

The implication runs deep. Every co-author carries liability for AI-related issues anywhere in the manuscript. The sections they personally wrote do not define the limits of their accountability. If a literature review co-authored by a postdoc contains a hallucinated citation, every name on the by-line shares responsibility. The corresponding author handles the administrative duty, but other authors do not get a pass on misconduct findings.

This is the part most research teams have not absorbed. Disclosure is not a courtesy from the section's author. It is a precondition for every co-author's signature on the final manuscript.

The Four Collaborative Scenarios Where Disclosure Breaks Down

Four common situations lead to incomplete or contested disclosure on multi-author papers.

Scenario 1: The unknown contributor. A junior researcher uses AI heavily for literature synthesis but never tells the team because they fear being seen as less competent. The senior authors approve the manuscript without knowing AI was involved. Disclosure ends up incomplete by accident.

Scenario 2: The minimizer. A co-author used AI for "just grammar" or "just polishing" and decides this does not warrant disclosure. The line between assistive editing (often exempt) and generative editing (usually requiring disclosure) is genuinely ambiguous. The safer default is to disclose. Most disputes within research teams happen at this fault line.

Scenario 3: The refusal. One co-author insists on disclosing their AI use. Another insists their use was trivial and refuses to declare it. The corresponding author gets stuck choosing between an incomplete disclosure or removing a co-author over a procedural disagreement.

Scenario 4: The discovery after submission. The manuscript is under review when one co-author admits to AI assistance that was not in the disclosure statement. The team faces a hard choice: update the journal mid-review and risk the perception of concealment, or wait until acceptance and risk a more serious integrity finding.

Each of these has the same root cause. AI use was an individual decision, never a team conversation.

Why This is Now a Retraction Risk, not a Disclosure Problem

In August 2025, the Committee on Publication Ethics (COPE) updated its retraction guidelines to include "undisclosed involvement of artificial intelligence" as grounds for retraction, alongside paper mills, identity theft, and fraud. This is a structural shift. Failure to disclose AI is treated as a form of misrepresentation that justifies pulling a published paper.

The collaborative implication is sharp. If any single co-author's undisclosed AI use surfaces post-publication, the entire paper is at risk of retraction. Other co-authors do not get a partial reprieve because the AI was used in a section they did not write. They share the integrity finding.

A 2025 American Association for Cancer Research analysis reported in Science found that 36% of submitted abstracts contained AI-generated text. Only 9% of authors disclosed AI use. Across multi-author submissions, the disclosure gap is almost certainly wider, because each additional author adds another point at which AI use slips below team-level visibility.

A Pre-submission Protocol for Collaborative Teams

The fix is procedural, not technological. Research teams that treat AI disclosure as a team-level decision rather than an individual one avoid the breakdowns described above. The following four-step protocol takes about an hour to implement on a typical multi-author manuscript.

Step 1: Have the conversation at project kickoff, not at submission. Before drafting begins, the corresponding author should send a short message asking each co-author which AI tools they expect to use, for which tasks, and in which sections. The Slade et al. study suggests fewer than half of teams currently do this. Making it routine eliminates most downstream disputes.

Step 2: Maintain a shared AI use log. Use a single shared document, accessible to all co-authors, where each contributor records the tool name and version, the specific task (drafting, summarizing, code generation, language editing), the section or content area, and the date. This log becomes the source of truth for the disclosure statement.

Step 3: Have the corresponding author draft a single consolidated disclosure statement. Rather than collecting individual disclosures, the corresponding author synthesizes the log into one statement. A working template:

Multiple authors used [list of tools and versions] during manuscript preparation. Specifically, [Author A] used [tool] for [task] in [section], [Author B] used [tool] for [task] in [section]. All AI-generated or AI-modified content was reviewed and verified by the human authors, who take full responsibility for the integrity and accuracy of the manuscript.

Step 4: Require explicit sign-off from every co-author on the final disclosure. Every author should confirm in writing that the consolidated disclosure accurately reflects their AI use. This converts an individual silence into a collective declaration. It also protects every co-author from being held responsible for someone else's omission.

For research groups looking to produce these statements consistently across projects, Enago's AI Disclosure Statement Generator walks teams through publisher-aligned fields and produces a disclosure ready for ICMJE-compliant manuscripts.

What PIs and Senior Authors Should do this Quarter

The collaborative AI disclosure problem will not solve itself. Three specific actions move a research group from individual silence to team-level transparency.

First, add AI use to the standard authorship agreement template used at project initiation. Most research groups already use a written authorship agreement for collaborative work. AI disclosure belongs in it as a standing item, like data ownership and authorship order.

Second, treat undisclosed AI use as a research integrity issue at the institutional level, with the same weight given to other forms of misconduct. Departments and graduate programs should establish that hidden AI use within a collaborative manuscript falls within research misconduct policy.

Third, build AI literacy across research teams at every level, from graduate students to senior PIs. The Kentucky study found strong interest in self-guided training, particularly on data security, ethical use, and team integration. Institutions that invest in this training now will face fewer disclosure crises later. The Enago Responsible AI Initiative provides resources and structured guidance for institutions building this kind of capacity.

The Conversation That Needs to Happen

Single-author disclosure tests memory. Collaborative disclosure tests trust. When co-authors do not know what AI their colleagues used, the disclosure statement becomes an estimate at best and a fiction at worst.

Publishers have already shifted their position. Undisclosed AI use is grounds for retraction. Funders have shifted too. NIH and NSF have explicit policies on AI use in applications and review. The remaining gap sits at the research team level, where AI is still treated as a personal tool rather than a shared methodological choice. Closing the gap is not a technical problem. It requires the simplest and hardest thing in collaborative research: a direct conversation, early, before anyone has typed a prompt.